October 2016
PREPARING FOR THE FUTURE OF ARTIFICIAL INTELLIGENCE
National Science and Technology Council
PREPARING FOR THE FUTURE OF ARTIFICIAL INTELLIGENCE
Executive Office of the President
National Science and Technology Council
Committee on Technology
About the National Science and Technology Council
The National Science and Technology Council (NSTC) is the principal means by which the Executive Branch coordinates science and technology policy across the diverse entities that make up the Federal research and development (R&D) enterprise. One of the NSTC’s primary objectives is establishing clear national goals for Federal science and technology investments. The NSTC prepares R&D packages aimed at accomplishing multiple national goals. The NSTC’s work is organized under five committees:
Environment, Natural Resources, and Sustainability; Homeland and National Security; Science,
Technology, Engineering, and Mathematics (STEM) Education; Science; and Technology. Each of these committees oversees subcommittees and working groups that are focused on different aspects of science and technology. More information is available at www.whitehouse.gov/ostp/nstc.
About the Office of Science and Technology Policy
The Office of Science and Technology Policy (OSTP) was established by the National Science and Technology Policy, Organization, and Priorities Act of 1976. OSTP’s responsibilities include advising the President in policy formulation and budget development on questions in which science and technology are important elements; articulating the President’s science and technology policy and programs; and fostering strong partnerships among Federal, state, and local governments, and the scientific communities in industry and academia. The Director of OSTP also serves as Assistant to the President for Science and Technology and manages the NSTC. More information is available at www.whitehouse.gov/ostp.
Acknowledgments
This document was developed through the contributions of staff from OSTP, other components of the Executive Office of the President, and other departments and agencies. A special thanks and appreciation to everyone who contributed.
Copyright Information
This is a work of the U.S. Government and is in the public domain. It may be freely distributed, copied, and translated; acknowledgment of publication by the Office of Science and Technology Policy is appreciated. Any translation should include a disclaimer that the accuracy of the translation is the responsibility of the translator and not OSTP. It is requested that a copy of any translation be sent to OSTP. This work is available for worldwide use and reuse and under the Creative Commons CC0 1.0 Universal license.
EXECUTIVE OFFICE OF THE PRESIDENT NATIONAL SCIENCE AND TECHNOLOGY COUNCIL
WASHINGTON, D.C. 20502
October 12, 2016
Dear colleagues:
Advances in Artificial Intelligence (AI) technology have opened up new markets and new opportunities for progress in critical areas such as health, education, energy, and the environment. In recent years, machines have surpassed humans in the performance of certain specific tasks, such as some aspects of image recognition. Experts forecast that rapid progress in the field of specialized artificial intelligence will continue. Although it is very unlikely that machines will exhibit broadly-applicable intelligence comparable to or exceeding that of humans in the next 20 years, it is to be expected that machines will reach and exceed human performance on more and more tasks.
As a contribution toward preparing the United States for a future in which AI plays a growing role, this report surveys the current state of AI, its existing and potential applications, and the questions that are raised for society and public policy by progress in AI. The report also makes recommendations for specific further actions by Federal agencies and other actors. A companion document lays out a strategic plan for Federally-funded research and development in AI. Additionally, in the coming months, the Administration will release a follow-on report exploring in greater depth the effect of AI-driven automation on jobs and the economy.
The report was developed by the NSTC’s Subcommittee on Machine Learning and Artificial Intelligence, which was chartered in May 2016 to foster interagency coordination, to provide technical and policy advice on topics related to AI, and to monitor the development of AI technologies across industry, the research community, and the Federal Government. The report was reviewed by the NSTC Committee on Technology, which concurred with its contents. The report follows a series of public-outreach activities spearheaded by the White House Office of Science and Technology Policy (OSTP) in 2016, which included five public workshops co-hosted with universities and other associations that are referenced in this report.
OSTP also published a Request for Information (RFI) in June 2016, which received 161 responses. The submitted comments were published by OSTP on September 6, 2016. Consistent with the role of Big Data as an enabler of AI, this report builds on three previous Administration reports on Big Data referenced in this report.
In the coming years, AI will continue to contribute to economic growth and will be a valuable tool for improving the world, as long as industry, civil society, and government work together to develop the positive aspects of the technology, manage its risks and challenges, and ensure that everyone has the opportunity to help in building an AI-enhanced society and to participate in its benefits.
Sincerely,
John P. Holdren Megan Smith
Assistant to the President for Science and Technology U.S. Chief Technology Officer Director, Office of Science and Technology Policy
National Science and Technology Council Chair
John P. Holdren
Assistant to the President for Science and Technology and Director, Office of Science and Technology Policy
Staff Afua Bruce Executive Director
Office of Science and Technology Policy
Subcommittee on
Machine Learning and Artificial Intelligence Co-Chair
Ed Felten
Deputy U.S. Chief Technology Officer Office of Science and Technology Policy Co-Chair
Executive Secretary Terah Lyons
Policy Advisor to the U.S. Chief Technology Officer
Office of Science and Technology Policy Michael Garris
Senior Scientist
National Institute of Standards and Technology U.S. Department of Commerce
The following Federal departments and agencies are represented on the Subcommittee on Machine Learning and Artificial Intelligence and through it, work together to monitor the state of the art in
machine learning (ML) and AI (within the Federal Government, in the private sector, and internationally), to watch for the arrival of important technology milestones in the development of AI, to coordinate the use of and foster the sharing of knowledge and best practices about ML and AI by the Federal
Government, and to consult in the development of Federal research and development priorities in AI:
Department of Commerce (Co-Chair) Department of Defense
Department of Education Department of Energy
Department of Health and Human Services Department of Homeland Security
Department of Justice Department of Labor Department of State
Department of Transportation Department of Treasury
Department of Veterans Affairs United States Agency for International Development
Central Intelligence Agency General Services Administration National Science Foundation National Security Agency
National Aeronautics and Space Administration Office of the Director of National Intelligence Social Security Administration
The following offices of the Executive Office of the President are also represented on the Subcommittee:
Council of Economic Advisers Domestic Policy Council
Office of Management and Budget
Office of Science and Technology Policy (Co- Chair)
Office of the Vice President National Economic Council National Security Council
Contents
Executive Summary ... 1
Introduction ... 5
A Brief History of AI ... 5
What is Artificial Intelligence? ... 6
The Current State of AI ... 7
Public Outreach and Development of this Report ... 12
Applications of AI for Public Good ... 13
AI in the Federal Government ... 15
AI and Regulation ... 17
Case Study: Autonomous Vehicles and Aircraft ... 18
Research and Workforce ... 23
Monitoring Progress in AI... 23
Federal Support for AI Research ... 25
Workforce Development and Diversity ... 26
AI, Automation, and the Economy ... 29
Fairness, Safety, and Governance ... 30
Justice, Fairness, and Accountability ... 30
Safety and Control... 32
Global Considerations and Security ... 35
International Cooperation ... 35
AI and Cybersecurity ... 36
AI in Weapon Systems ... 37
Conclusion ... 39
Recommendations in this Report ... 40
Acronyms... 43
References ... 45
1
Executive Summary
As a contribution toward preparing the United States for a future in which Artificial Intelligence (AI) plays a growing role, we survey the current state of AI, its existing and potential applications, and the questions that are raised for society and public policy by progress in AI. We also make recommendations for specific further actions by Federal agencies and other actors. A companion document called the National Artificial Intelligence Research and Development Strategic Plan lays out a strategic plan for Federally-funded research and development in AI.
Applications of AI for Public Good
One area of great optimism about AI and machine learning is their potential to improve people’s lives by helping to solve some of the world’s greatest challenges and inefficiencies. Many have compared the promise of AI to the transformative impacts of advancements in mobile computing. Public- and private- sector investments in basic and applied R&D on AI have already begun reaping major benefits to the public in fields as diverse as health care, transportation, the environment, criminal justice, and economic inclusion. The effectiveness of government itself is being increased as agencies build their capacity to use AI to carry out their missions more quickly, responsively, and efficiently.
AI and Regulation
AI has applications in many products, such as cars and aircraft, which are subject to regulation designed to protect the public from harm and ensure fairness in economic competition. How will the incorporation of AI into these products affect the relevant regulatory approaches? In general, the approach to regulation of AI-enabled products to protect public safety should be informed by assessment of the aspects of risk that the addition of AI may reduce alongside the aspects of risk that it may increase. If a risk falls within the bounds of an existing regulatory regime, moreover, the policy discussion should start by considering whether the existing regulations already adequately address the risk, or whether they need to be adapted to the addition of AI. Also, where regulatory responses to the addition of AI threaten to increase the cost of compliance, or slow the development or adoption of beneficial innovations, policymakers should consider how those responses could be adjusted to lower costs and barriers to innovation without adversely impacting safety or market fairness.
Currently relevant examples of the regulatory challenges that AI-enabled products present are found in the cases of automated vehicles (AVs, such as self-driving cars) and AI-equipped unmanned aircraft systems (UAS, or “drones”). In the long run, AVs will likely save many lives by reducing driver error and increasing personal mobility, and UAS will offer many economic benefits. Yet public safety must be protected as these technologies are tested and begin to mature. The Department of Transportation (DOT) is using an approach to evolving the relevant regulations that is based on building expertise in the Department, creating safe spaces and test beds for experimentation, and working with industry and civil society to evolve performance-based regulations that will enable more uses as evidence of safe operation accumulates.
Research and Workforce
Government also has an important role to play in the advancement of AI through research and development and the growth of a skilled, diverse workforce. A separate strategic plan for Federally- funded AI research and development is being released in conjunction with this report. The plan discusses the role of Federal R&D, identifies areas of opportunity, and recommends ways to coordinate R&D to maximize benefit and build a highly-trained workforce.
2
Given the strategic importance of AI, moreover, it is appropriate for the Federal Government to monitor developments in the field worldwide in order to get early warning of important changes arising elsewhere in case these require changes in U.S. policy.
The rapid growth of AI has dramatically increased the need for people with relevant skills to support and advance the field. An AI-enabled world demands a data-literate citizenry that is able to read, use,
interpret, and communicate about data, and participate in policy debates about matters affected by AI. AI knowledge and education are increasingly emphasized in Federal Science, Technology, Engineering, and Mathematics (STEM) education programs. AI education is also a component of Computer Science for All, the President’s initiative to empower all American students from kindergarten through high school to learn computer science and be equipped with the computational thinking skills they need in a technology- driven world.
Economic Impacts of AI
AI’s central economic effect in the short term will be the automation of tasks that could not be automated before. This will likely increase productivity and create wealth, but it may also affect particular types of jobs in different ways, reducing demand for certain skills that can be automated while increasing demand for other skills that are complementary to AI. Analysis by the White House Council of Economic
Advisors (CEA) suggests that the negative effect of automation will be greatest on lower-wage jobs, and that there is a risk that AI-driven automation will increase the wage gap between less-educated and more- educated workers, potentially increasing economic inequality. Public policy can address these risks, ensuring that workers are retrained and able to succeed in occupations that are complementary to, rather than competing with, automation. Public policy can also ensure that the economic benefits created by AI are shared broadly, and assure that AI responsibly ushers in a new age in the global economy.
Fairness, Safety, and Governance
As AI technologies move toward broader deployment, technical experts, policy analysts, and ethicists have raised concerns about unintended consequences of widespread adoption. Use of AI to make consequential decisions about people, often replacing decisions made by human-driven bureaucratic processes, leads to concerns about how to ensure justice, fairness, and accountability—the same concerns voiced previously in the Administration’s Big Data: Seizing Opportunities, Preserving Values report of 2014,1 as well as the Report to the President on Big Data and Privacy: A Technological Perspective published by the President’s Council of Advisors on Science and Technology in 2014.2 Transparency concerns focus not only on the data and algorithms involved, but also on the potential to have some form of explanation for any AI-based determination. Yet AI experts have cautioned that there are inherent challenges in trying to understand and predict the behavior of advanced AI systems.
Use of AI to control physical-world equipment leads to concerns about safety, especially as systems are exposed to the full complexity of the human environment. A major challenge in AI safety is building systems that can safely transition from the “closed world” of the laboratory into the outside “open world”
where unpredictable things can happen. Adapting gracefully to unforeseen situations is difficult yet necessary for safe operation. Experience in building other types of safety-critical systems and
infrastructure, such as aircraft, power plants, bridges, and vehicles, has much to teach AI practitioners
1 “Big Data: Seizing Opportunities, Preserving Values,” Executive Office of the President, May 2014, https://www.whitehouse.gov/sites/default/files/docs/big_data_privacy_report_may_1_2014.pdf.
2 The President’s Council of Advisors on Science and Technology, “Report to the President: Big Data and Privacy:
A Technological Perspective,” Executive Office of the President, May 2014,
https://www.whitehouse.gov/sites/default/files/microsites/ostp/PCAST/pcast_big_data_and_privacy_- _may_2014.pdf.
3
about verification and validation, how to build a safety case for a technology, how to manage risk, and how to communicate with stakeholders about risk.
At a technical level, the challenges of fairness and safety are related. In both cases, practitioners strive to avoid unintended behavior, and to generate the evidence needed to give stakeholders justified confidence that unintended failures are unlikely.
Ethical training for AI practitioners and students is a necessary part of the solution. Ideally, every student learning AI, computer science, or data science would be exposed to curriculum and discussion on related ethics and security topics. However, ethics alone is not sufficient. Ethics can help practitioners understand their responsibilities to all stakeholders, but ethical training should be augmented with technical tools and methods for putting good intentions into practice by doing the technical work needed to prevent
unacceptable outcomes.
Global Considerations and Security
AI poses policy questions across a range of areas in international relations and security. AI has been a topic of interest in recent international discussions as countries, multilateral institutions, and other stakeholders have begun to access the benefits and challenges of AI. Dialogue and cooperation between these entities could help advance AI R&D and harness AI for good, while also addressing shared challenges.
Today’s AI has important applications in cybersecurity, and is expected to play an increasing role for both defensive and offensive cyber measures. Currently, designing and operating secure systems requires significant time and attention from experts. Automating this expert work partially or entirely may increase security across a much broader range of systems and applications at dramatically lower cost, and could increase the agility of the Nation’s cyber-defenses. Using AI may help maintain the rapid response required to detect and react to the landscape of evolving threats.
Challenging issues are raised by the potential use of AI in weapon systems. The United States has
incorporated autonomy in certain weapon systems for decades, allowing for greater precision in the use of weapons and safer, more humane military operations. Nonetheless, moving away from direct human control of weapon systems involves some risks and can raise legal and ethical questions.
The key to incorporating autonomous and semi-autonomous weapon systems into American defense planning is to ensure that U.S. Government entities are always acting in accordance with international humanitarian law, taking appropriate steps to control proliferation, and working with partners and Allies to develop standards related to the development and use of such weapon systems. The United States has actively participated in ongoing international discussion on Lethal Autonomous Weapon Systems, and anticipates continued robust international discussion of these potential weapon systems. Agencies across the U.S. Government are working to develop a single, government-wide policy, consistent with
international humanitarian law, on autonomous and semi-autonomous weapons.
Preparing for the Future
AI holds the potential to be a major driver of economic growth and social progress, if industry, civil society, government, and the public work together to support development of the technology with thoughtful attention to its potential and to managing its risks.
The U.S. Government has several roles to play. It can convene conversations about important issues and help to set the agenda for public debate. It can monitor the safety and fairness of applications as they develop, and adapt regulatory frameworks to encourage innovation while protecting the public. It can provide public policy tools to ensure that disruption in the means and methods of work enabled by AI increases productivity while avoiding negative economic consequences for certain sectors of the workforce. It can support basic research and the application of AI to public good. It can support
development of a skilled, diverse workforce. And government can use AI itself to serve the public faster,
4
more effectively, and at lower cost. Many areas of public policy, from education and the economic safety net, to defense, environmental preservation, and criminal justice, will see new opportunities and new challenges driven by the continued progress of AI. The U.S. Government must continue to build its capacity to understand and adapt to these changes.
As the technology of AI continues to develop, practitioners must ensure that AI-enabled systems are governable; that they are open, transparent, and understandable; that they can work effectively with people; and that their operation will remain consistent with human values and aspirations. Researchers and practitioners have increased their attention to these challenges, and should continue to focus on them.
Developing and studying machine intelligence can help us better understand and appreciate our human intelligence. Used thoughtfully, AI can augment our intelligence, helping us chart a better and wiser path forward.
A full list of the recommendations made in this report is on page 40.
5
Introduction
Artificial Intelligence (AI) has the potential to help address some of the biggest challenges that society faces. Smart vehicles may save hundreds of thousands of lives every year worldwide, and increase mobility for the elderly and those with disabilities. Smart buildings may save energy and reduce carbon emissions. Precision medicine may extend life and increase quality of life. Smarter government may serve citizens more quickly and precisely, better protect those at risk, and save money. AI-enhanced education may help teachers give every child an education that opens doors to a secure and fulfilling life. These are just a few of the potential benefits if the technology is developed with an eye to its benefits and with careful consideration of its risks and challenges.
The United States has been at the forefront of foundational research in AI, primarily supported for most of the field’s history by Federal research funding and work at government laboratories. The Federal
Government’s support for unclassified AI R&D is managed through the Networking and Information Technology Research and Development (NITRD) program, and supported primarily by the Defense Advanced Research Projects Agency (DARPA), the National Science Foundation (NSF), the National Institutes of Health (NIH), the Office of Naval Research (ONR), and the Intelligence Advanced Research Projects Activity (IARPA). Major national research efforts such as the National Strategic Computing Initiative, the Big Data Initiative, and the Brain Research through Advancing Innovative
Neurotechnologies (BRAIN) Initiative also contribute indirectly to the progress of AI research. The current and projected benefits of AI technology are large, adding to the Nation’s economic vitality and to the productivity and well-being of its people. A companion document lays out a strategic plan for Federally-funded research and development in AI.
As a contribution toward preparing the United States for a future in which AI plays a growing role, we survey the current state of AI, its existing and potential applications, and the questions that progress in AI raise for society and public policy. We also make recommendations for specific further actions by Federal agencies and other actors.
A Brief History of AI
Endowing computers with human-like intelligence has been a dream of computer experts since the dawn of electronic computing. Although the term “Artificial Intelligence” was not coined until 1956, the roots of the field go back to at least the 1940s,3 and the idea of AI was crystalized in Alan Turing’s famous 1950 paper, “Computing Machinery and Intelligence.” Turing’s paper posed the question: “Can machines think?” It also proposed a test for answering that question,4 and raised the possibility that a machine might be programmed to learn from experience much as a young child does.
In the ensuing decades, the field of AI went through ups and downs as some AI research problems proved more difficult than anticipated and others proved insurmountable with the technologies of the time. It wasn’t until the late 1990s that research progress in AI began to accelerate, as researchers focused more on sub-problems of AI and the application of AI to real-world problems such as image recognition and medical diagnosis. An early milestone was the 1997 victory of IBM’s chess-playing computer Deep Blue
3 See, e.g., Warren S. McCulloch and Walter H. Pitts, “A Logical Calculus of the Ideas Immanent in Nervous Activity,” Bulletin of Mathematical Biophysics, 5:115-133, 1943.
4 Restated in modern terms, the “Turing Test” puts a human judge in a text-based chat room with either another person or a computer. The human judge can interrogate the other party and carry on a conversation, and then the judge is asked to guess whether the other party is a person or a computer. If a computer can consistently fool human judges in this game, then the computer is deemed to be exhibiting intelligence.
6
over world champion Garry Kasparov. Other significant breakthroughs included DARPA’s Cognitive Agent that Learns and Organizes (CALO), which led to Apple Inc.’s Siri; IBM’s question-answering computer Watson’s victory in the TV game show “Jeopardy!”; and the surprising success of self-driving cars in the DARPA Grand Challenge competitions in the 2000s.
The current wave of progress and enthusiasm for AI began around 2010, driven by three factors that built upon each other: the availability of big data from sources including e-commerce, businesses, social media, science, and government; which provided raw material for dramatically improved machine learning approaches and algorithms; which in turn relied on the capabilities of more powerful
computers.5 During this period, the pace of improvement surprised AI experts. For example, on a popular image recognition challenge6 that has a 5 percent human error rate according to one error measure, the best AI result improved from a 26 percent error rate in 2011 to 3.5 percent in 2015.
Simultaneously, industry has been increasing its investment in AI. In 2016, Google Chief Executive Officer (CEO) Sundar Pichai said, “Machine learning [a subfield of AI] is a core, transformative way by which we’re rethinking how we’re doing everything. We are thoughtfully applying it across all our products, be it search, ads, YouTube, or Play. And we’re in early days, but you will see us — in a systematic way — apply machine learning in all these areas.”7 This view of AI broadly impacting how software is created and delivered was widely shared by CEOs in the technology industry, including Ginni Rometty of IBM, who has said that her organization is betting the company on AI.8
What is Artificial Intelligence?
There is no single definition of AI that is universally accepted by practitioners. Some define AI loosely as a computerized system that exhibits behavior that is commonly thought of as requiring intelligence.
Others define AI as a system capable of rationally solving complex problems or taking appropriate actions to achieve its goals in whatever real world circumstances it encounters.
Experts offer differing taxonomies of AI problems and solutions. A popular AI textbook9 used the following taxonomy: (1) systems that think like humans (e.g., cognitive architectures and neural networks); (2) systems that act like humans (e.g., pass the Turing test via natural language processing;
knowledge representation, automated reasoning, and learning), (3) systems that think rationally (e.g.,
5 A more detailed history of AI is available in the Appendix of the AI 100 Report. Peter Stone, Rodney Brooks, Erik Brynjolfsson, Ryan Calo, Oren Etzioni, Greg Hager, Julia Hirschberg, Shivaram Kalyanakrishnan, Ece Kamar, Sarit Kraus, Kevin Leyton-Brown, David Parkes, William Press, AnnaLee Saxenian, Julie Shah, Milind Tambe, and Astro Teller, "Artificial Intelligence and Life in 2030," One Hundred Year Study on Artificial Intelligence: Report of the 2015-2016 Study Panel, Stanford University, Stanford, CA, September 2016, http://ai100.stanford.edu/2016- report.
6 The ImageNet Large Scale Visual Recognition Challenge provides a set of photographic images and asks for an accurate description of what is depicted in each image. Statistics in the text refer to the “classification error” metric in the “classification+localization with provided training data” task. See http://image-net.org/challenges/LSVRC/.
7 Steven Levy, “How Google is Remaking Itself as a Machine Learning First Company,” Backchannel, June 22, 2016, https://backchannel.com/how-google-is-remaking-itself-as-a-machine-learning-first-company-ada63defcb70.
8 See, e.g., Andrew Nusca, “IBM’s CEO Thinks Every Digital Business Will Become a Cognitive Computing Business,” Fortune, June 1 2016. (“[IBM] CEO Ginni Rometty is optimistic that the company’s wager on ‘cognitive computing,’ the term it uses for applied artificial intelligence and machine learning technologies, is the biggest bet the company will make in its 105-year history.”)
9 Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (3rd Edition) (Essex, England:
Pearson, 2009).
7
logic solvers, inference, and optimization); and (4) systems that act rationally (e.g., intelligent software agents and embodied robots that achieve goals via perception, planning, reasoning, learning,
communicating, decision-making, and acting). Separately, venture capitalist Frank Chen broke down the problem space of AI into five general categories: logical reasoning, knowledge representation, planning and navigation, natural language processing, and perception.10 And AI researcher Pedro Domingos ascribed AI researchers to five “tribes” based on the methods they use: “symbolists” use logical reasoning based on abstract symbols, “connectionists” build structures inspired by the human brain;
“evolutionaries” use methods inspired by Darwinian evolution; “Bayesians” use probabilistic inference;
and “analogizers” extrapolate from similar cases seen previously.11
This diversity of AI problems and solutions, and the foundation of AI in human evaluation of the performance and accuracy of algorithms, makes it difficult to clearly define a bright-line distinction between what constitutes AI and what does not. For example, many techniques used to analyze large volumes of data were developed by AI researchers and are now identified as “Big Data” algorithms and systems. In some cases, opinion may shift, meaning that a problem is considered as requiring AI before it has been solved, but once a solution is well known it is considered routine data processing. Although the boundaries of AI can be uncertain and have tended to shift over time, what is important is that a core objective of AI research and applications over the years has been to automate or replicate intelligent behavior.
The Current State of AI
Remarkable progress has been made on what is known as Narrow AI, which addresses specific
application areas such as playing strategic games, language translation, self-driving vehicles, and image recognition.12 Narrow AI underpins many commercial services such as trip planning, shopper
recommendation systems, and ad targeting, and is finding important applications in medical diagnosis, education, and scientific research. These have all had significant societal benefits and have contributed to the economic vitality of the Nation.13
General AI (sometimes called Artificial General Intelligence, or AGI) refers to a notional future AI system that exhibits apparently intelligent behavior at least as advanced as a person across the full range of cognitive tasks. A broad chasm seems to separate today’s Narrow AI from the much more difficult challenge of General AI. Attempts to reach General AI by expanding Narrow AI solutions have made little headway over many decades of research. The current consensus of the private-sector expert community, with which the NSTC Committee on Technology concurs, is that General AI will not be achieved for at least decades.14
10 Frank Chen, “AI, Deep Learning, and Machine Learning: A Primer,” Andreessen Horowitz, June 10, 2016, http://a16z.com/2016/06/10/ai-deep-learning-machines.
11 Pedro Domingos, The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World (New York, New York: Basic Books, 2015).
12 Narrow AI is not a single technical approach, but rather a set of discrete problems whose solutions rely on a toolkit of AI methods along with some problem-specific algorithms. The diversity of Narrow AI problems and solutions, and the apparent need to develop specific methods for each Narrow AI application, has made it infeasible to “generalize” a single Narrow AI solution to produce intelligent behavior of general applicability.
13 Mike Purdy and Paul Daugherty, “Why Artificial Intelligence is the Future of Growth,” Accenture, 2016, https://www.accenture.com/us-en/_acnmedia/PDF-33/Accenture-Why-AI-is-the-Future-of-Growth.pdf.
14 Expert opinion on the expected arrival date of AGI ranges from 2030 to centuries from now. There is a long history of excessive optimism about AI. For example, AI pioneer Herb Simon predicted in 1957 that computers
8
People have long speculated on the implications of computers becoming more intelligent than humans.
Some predict that a sufficiently intelligent AI could be tasked with developing even better, more intelligent systems, and that these in turn could be used to create systems with yet greater intelligence, and so on, leading in principle to an “intelligence explosion” or “singularity” in which machines quickly race far ahead of humans in intelligence.15
In a dystopian vision of this process, these super-intelligent machines would exceed the ability of
humanity to understand or control. If computers could exert control over many critical systems, the result could be havoc, with humans no longer in control of their destiny at best and extinct at worst. This scenario has long been the subject of science fiction stories, and recent pronouncements from some influential industry leaders have highlighted these fears.
A more positive view of the future held by many researchers sees instead the development of intelligent systems that work well as helpers, assistants, trainers, and teammates of humans, and are designed to operate safely and ethically.
The NSTC Committee on Technology’s assessment is that long-term concerns about super-intelligent General AI should have little impact on current policy. The policies the Federal Government should adopt in the near-to-medium term if these fears are justified are almost exactly the same policies the Federal Government should adopt if they are not justified. The best way to build capacity for addressing the longer-term speculative risks is to attack the less extreme risks already seen today, such as current security, privacy, and safety risks, while investing in research on longer-term capabilities and how their challenges might be managed. Additionally, as research and applications in the field continue to mature, practitioners of AI in government and business should approach advances with appropriate consideration of the long-term societal and ethical questions – in additional to just the technical questions – that such advances portend. Although prudence dictates some attention to the possibility that harmful super- intelligence might someday become possible, these concerns should not be the main driver of public policy for AI.
Machine Learning
Machine learning is one of the most important technical approaches to AI and the basis of many recent advances and commercial applications of AI. Modern machine learning is a statistical process that starts with a body of data and tries to derive a rule or procedure that explains the data or can predict future data.
This approach—learning from data—contrasts with the older “expert system” approach to AI, in which programmers sit down with human domain experts to learn the rules and criteria used to make decisions, and translate those rules into software code. An expert system aims to emulate the principles used by human experts, whereas machine learning relies on statistical methods to find a decision procedure that works well in practice.
An advantage of machine learning is that it can be used even in cases where it is infeasible or difficult to write down explicit rules to solve a problem. For example, a company that runs an online service might use machine learning to detect user log-in attempts that are fraudulent. The company might start with a large data set of past login attempts, with each attempt labeled as fraudulent or not using the benefit of
would outplay humans at chess within a decade, an outcome that required 40 years to occur. Early predictions about automated language translation also proved wildly optimistic, with the technology only becoming usable (and by no means fully fluent) in the last several years. It is tempting but incorrect to extrapolate from the ability to solve one particular task to imagine machines with a much broader and deeper range of capabilities and to overlook the huge gap between narrow task-oriented performance and the type of general intelligence that people exhibit.
15 It is far from certain that this sort of explosive growth in intelligence is likely, or even possible. Another plausible extrapolation from current knowledge is that machine intelligence will continue to increase gradually even after surpassing human intelligence.
9
hindsight. Based on this data set, the company could use machine learning to derive a rule to apply to future login attempts that predicts which attempts are more likely to be fraudulent and should be subjected to extra security measures. In a sense, machine learning is not an algorithm for solving a specific problem, but rather a more general approach to finding solutions for many different problems, given data about them.
To apply machine learning, a practitioner starts with a historical data set, which the practitioner divides into a training set and a test set. The practitioner chooses a model, or mathematical structure that
characterizes a range of possible decision-making rules with adjustable parameters. A common analogy is that the model is a “box” that applies a rule, and the parameters are adjustable knobs on the front of the box that control how the box operates. In practice, a model might have many millions of parameters.
The practitioner also defines an objective function used to evaluate the desirability of the outcome that results from a particular choice of parameters. The objective function will typically contain parts that reward the model for closely matching the training set, as well as parts that reward the use of simpler rules.
Training the model is the process of adjusting the parameters to maximize the objective function.
Training is the difficult technical step in machine learning. A model with millions of parameters will have astronomically more possible outcomes than any algorithm could ever hope to try, so successful training algorithms have to be clever in how they explore the space of parameter settings so as to find very good settings with a feasible level of computational effort.
Once a model has been trained, the practitioner can use the test set to evaluate the accuracy and effectiveness of the model. The goal of machine learning is to create a trained model that will
generalize—it will be accurate not only on examples in the training set, but also on future cases that it has never seen before. While many of these models can achieve better-than-human performance on narrow tasks such as image labeling, even the best models can fail in unpredictable ways. For example, for many image labeling models it is possible to create images that clearly appear to be random noise to a human but will be falsely labeled as a specific object with high confidence by a trained model.16
Another challenge in using machine learning is that it is typically not possible to extract or generate a straightforward explanation for why a particular trained model is effective. Because trained models have a very large number of adjustable parameters—often hundreds of millions or more—training may yield a model that "works," in the sense of matching the data, but is not necessarily the simplest model that works. In human decision-making, any opacity in the process is typically due to not having enough information about why a decision was reached, because the decider may be unable to articulate why the decision “felt right.” With machine learning, everything about the decision procedure is known with mathematical precision, but there may be simply too much information to interpret clearly.
Deep Learning
In recent years, some of the most impressive advancements in machine learning have been in the subfield of deep learning, also known as deep network learning. Deep learning uses structures loosely inspired by the human brain, consisting of a set of units (or “neurons”). Each unit combines a set of input values to produce an output value, which in turn is passed on to other neurons downstream. For example, in an image recognition application, a first layer of units might combine the raw data of the image to recognize simple patterns in the image; a second layer of units might combine the results of the first layer to recognize patterns-of-patterns; a third layer might combine the results of the second layer; and so on.
16 See, e.g., Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy, “Explaining and Harnessing Adversarial Examples,” http://arxiv.org/pdf/1412.6572.pdf.
10
Deep learning networks typically use many layers—sometimes more than 100— and often use a large number of units at each layer, to enable the recognition of extremely complex, precise patterns in data.
In recent years, new theories of how to construct and train deep networks have emerged, as have larger, faster computer systems, enabling the use of much larger deep learning networks. The dramatic success of these very large networks at many machine learning tasks has come as a surprise to some experts, and is the main cause of the current wave of enthusiasm for machine learning among AI researchers and practitioners.
Autonomy and Automation
AI is often applied to systems that can control physical actuators or trigger online actions. When AI comes into contact with the everyday world, issues of autonomy, automation, and human-machine teaming arise.
Autonomy refers to the ability of a system to operate and adapt to changing circumstances with reduced or without human control. For example, an autonomous car could drive itself to its destination. Despite the focus in much of the literature on cars and aircraft, autonomy is a much broader concept that includes scenarios such as automated financial trading and automated content curation systems. Autonomy also includes systems that can diagnose and repair faults in their own operation, such as identifying and fixing security vulnerabilities.
Automation occurs when a machine does work that might previously have been done by a person.17 The term relates to both physical work and mental or cognitive work that might be replaced by AI.
Automation, and its impact on employment, have been significant social and economic phenomena since at least the Industrial Revolution. It is widely accepted that AI will automate some jobs, but there is more debate about whether this is just the next chapter in the history of automation or whether AI will affect the economy differently than past waves of automation have previously.
Human-Machine Teaming
In contrast to automation, where a machine substitutes for human work, in some cases a machine will complement human work. This may happen as a side-effect of AI development, or a system might be developed specifically with the goal of creating a human-machine team. Systems that aim to complement human cognitive capabilities are sometimes referred to as intelligence augmentation.
In many applications, a human-machine team can be more effective than either one alone, using the strengths of one to compensate for the weaknesses of the other. One example is in chess playing, where a weaker computer can often beat a stronger computer player, if the weaker computer is given a human teammate—this is true even though top computers are much stronger players than any human.18 Another example is in radiology. In one recent study, given images of lymph node cells, and asked to determine
17 Different definitions of “automation” are used in different settings. The definition used in the main text, involving the substitution of machine labor for human labor, is commonly used in economics. Another definition is used in the systems analysis setting in the Department of Defense (DoD): Automation means that the system functions with little or no human operator involvement. However the system performance is limited to the specific pre-programmed actions it has been designed to execute. Once the system is initiated by a human operator, it executes its task
according to those instructions and subroutines, which have been tested and validated. Typically these are well- defined tasks that have predetermined responses, i.e., rule-based responses in reasonably well-known and structured environments.
18 Garry Kasparov, “The Chess Master and the Computer,” New York Review of Books, February 11, 2010.
http://www.nybooks.com/articles/2010/02/11/the-chess-master-and-the-computer.
11
whether or not the cells contained cancer, an AI-based approach had a 7.5 percent error rate, where a human pathologist had a 3.5 percent error rate; a combined approach, using both AI and human input, lowered the error rate to 0.5 percent, representing an 85 percent reduction in error.19
19 Dayong Wang, Aditya Khosla, Rishab Gargeya, Humayun Irshad, Andrew H. Beck, “Deep Learning for Identifying Metastatic Breast Cancer,” June 18, 2016, https://arxiv.org/pdf/1606.05718v1.pdf.
12
Public Outreach and Development of this Report
This report was developed by the NSTC’s Subcommittee on Machine Learning and Artificial Intelligence, which was chartered in May 2016 to foster interagency coordination and provide technical and policy advice on topics related to AI, and to monitor the development of AI technologies across industry, the research community, and the Federal Government. The report follows a series of public outreach activities led by OSTP, designed to allow government officials thinking about these topics to learn from experts and from the public. This public outreach on AI included five co-hosted public workshops, and a public Request for Information (RFI). The public workshops were:
AI, Law, and Governance (May 24, in Seattle, co-hosted by OSTP, the National Economic Council (NEC), and the University of Washington);
AI for Social Good (June 7, in Washington DC, co-hosted by OSTP, the Association for the Advancement of AI (AAAI) and the Computing Community Consortium (CCC));
Future of AI: Emerging Topics and Societal Benefit at the Global Entrepreneurship Summit (June 23, in Palo Alto, co-hosted by OSTP and Stanford University);
AI Technology, Safety, and Control (June 28, in Pittsburgh, co-hosted by OSTP and Carnegie Mellon University); and
Social and Economic Impacts of AI (July 7, in New York, co-hosted by OSTP, NEC, and New York University).
In conjunction with each of the five workshops, the private-sector co-hosts organized separate meetings or conference sessions which government staff attended. Total in-person attendance at the public events was more than 2,000 people, in addition to international online streaming audiences, which included more than 3,500 people for the Washington, DC workshop livestream alone.
OSTP also published a Request for Information (RFI) seeking public comment on the topics of the workshops. The RFI closed on July 22, 2016 and received 161 responses. Comments submitted in response to the public RFI were published by OSTP on September 6, 2016.20
20 Ed Felten and Terah Lyons, “Public Input and Next Steps on the Future of Artificial Intelligence,” Medium, September 6 2016, https://medium.com/@USCTO/public-input-and-next-steps-on-the-future-of-artificial- intelligence-458b82059fc3.
13
Applications of AI for Public Good
One area of great optimism about AI and machine learning is their potential to improve people’s lives by helping to solve some of the world’s greatest challenges and inefficiencies. The promise of AI has been compared to the transformative impacts of advances in mobile computing.21 Public- and private-sector investments in basic and applied R&D on AI have already begun reaping major benefits for the public in fields as diverse as health care, transportation, the environment, criminal justice, and economic
inclusion.22
At Walter Reed Medical Center, the Department of Veteran Affairs is using AI to better predict medical complications and improve treatment of severe combat wounds, leading to better patient outcomes, faster healing, and lower costs.23 The same general approach—predicting complications to enable preventive treatment—has also reduced hospital-acquired infections at Johns Hopkins University.24 Given the current transition to electronic health records, predictive analysis of health data may play a key role across many health domains like precision medicine and cancer research.
In transportation, AI-enabled smarter traffic management applications are reducing wait times, energy use, and emissions by as much as 25 percent in some places.25 Cities are now beginning to leverage the type of responsive dispatching and routing used by ride-hailing services, and linking it with scheduling and tracking software for public transportation to provide just-in-time access to public transportation that can often be faster, cheaper and, in many cases, more accessible to the public.
Some researchers are leveraging AI to improve animal migration tracking by using AI image
classification software to analyze tourist photos from public social media sites. The software can identify individual animals in the photos and build a database of their migration using the data and location stamps on the photos. At OSTP’s AI for Social Good workshop, researchers talked about building some of the largest available datasets to-date on the populations and migrations of whales and large African animals, and about launching a project to track “The Internet of Turtles” to gain new insights about sea life.26 Other speakers described uses of AI to optimize the patrol strategy of anti-poaching agents, and to design habitat preservation strategies to maximize the genetic diversity of endangered populations.
21 Frank Chen, “AI, Deep Learning, and Machine Learning: A Primer,” Andreessen Horowitz, June 10, 2016, http://a16z.com/2016/06/10/ai-deep-learning-machines.
22The potential benefits of increasing access to digital technologies are detailed further in the World Bank Group’s Digital Dividends report. (“World Development Report 2016: Digital Dividends,” The World Bank Group, 2016, http://documents.worldbank.org/curated/en/896971468194972881/pdf/102725-PUB-Replacement-PUBLIC.pdf.)
23 Eric Elster, “Surgical Critical Care Initiative: Bringing Precision Medicine to the Critically Ill,” presentation at AI for Social Good workshop, Washington, DC, June 7, 2016, http://cra.org/ccc/wp-
content/uploads/sites/2/2016/06/Eric-Elster-AI-slides-min.pdf.
24 Katharine E. Henry, David N. Hager, Peter J. Pronovost, and Suchi Saria, "A targeted real-time early warning score (TREWScore) for septic shock," Science Translational Medicine 7, no. 299 (2015): 299ra122-299ra122.
25 Stephen F. Smith, “Smart Infrastructure for Urban Mobility,” presentation at AI for Social Good workshop, Washington, DC, June 7, 2016, http://cra.org/ccc/wp-content/uploads/sites/2/2016/06/Stephen-Smith-AI-slides.pdf.
26 Aimee Leslie, Christine Hof, Diego Amorocho, Tanya Berger-Wolf, Jason Holmberg, Chuck Stewart, Stephen G.
Dunbar, and Claire Jea,, “The Internet of Turtles,” April 12, 2016,
https://www.researchgate.net/publication/301202821_The_Internet_of_Turtles.
14
Autonomous sailboats and watercraft are already patrolling the oceans carrying sophisticated sensor instruments, collecting data on changes in Arctic ice and sensitive ocean ecosystems in operations that would be too expensive or dangerous for crewed vessels. Autonomous watercraft may be much cheaper to operate than manned ships, and may someday be used for enhanced weather prediction, climate
monitoring, or policing illegal fishing.27
AI also has the potential to improve aspects of the criminal justice system, including crime reporting, policing, bail, sentencing, and parole decisions. The Administration is exploring how AI can responsibly benefit current initiatives such as Data Driven Justice and the Police Data Initiative that seek to provide law enforcement and the public with data that can better inform decision-making in the criminal justice system, while also taking care to minimize the possibility that AI might introduce bias or inaccuracies due to deficiencies in the available data.
Several U.S. academic institutions have launched initiatives to use AI to tackle economic and social challenges. For example, the University of Chicago created an academic program that uses data science and AI to address public challenges such as unemployment and school dropouts.28 The University of Southern California launched the Center for Artificial Intelligence in Society, an institute dedicated to studying how computational game theory, machine learning, automated planning and multi-agent reasoning techniques can help to solve socially relevant problems like homelessness. Meanwhile, researchers at Stanford University are using machine learning in efforts to address global poverty by using AI to analyze satellite images of likely poverty zones to identify where help is needed most.29 Many uses of AI for public good rely on the availability of data that can be used to train machine learning models and test the performance of AI systems. Agencies and organizations with data that can be released without implicating personal privacy or trade secrets can help to enable the development of AI by making those data available to researchers. Standardizing data schemas and formats can reduce the cost and difficulty of making new data sets useful.
27 John Markoff, “No Sailors Needed: Robot Sailboats Scout the Oceans for Data,” The New York Times, September 4, 2016.
28 “Data Science for Social Good,” University of Chicago, https://dssg.uchicago.edu/.
29 Neal Jean, Marshall Burke, Michael Xie, W. Matthew Davis, David B. Lobell, and Stefano Ermon. "Combining satellite imagery and machine learning to predict poverty." Science 353, no. 6301 (2016): 790-794.
Recommendation 1: Private and public institutions are encouraged to examine whether and how they can responsibly leverage AI and machine learning in ways that will benefit society. Social justice and public policy institutions that do not typically engage with advanced technologies and data science in their work should consider partnerships with AI researchers and practitioners that can help apply AI tactics to the broad social problems these institutions already address in other ways.
Recommendation 2: Federal agencies should prioritize open training data and open data standards in AI. The government should emphasize the release of datasets that enable the use of AI to address social challenges. Potential steps may include developing an “Open Data for AI” initiative with the objective of releasing a significant number of government data sets to accelerate AI research and galvanize the use of open data standards and best practices across government, academia, and the private sector.
15
AI in the Federal Government
The Administration is working to develop policies and internal practices that will maximize the economic and societal benefits of AI and promote innovation. These policies and practices may include:
investing in basic and applied research and development (R&D);
serving as an early customer for AI technologies and their applications;
supporting pilot projects and creating testbeds in real-world settings;
making data sets available to the public;
sponsoring incentive prizes;
identifying and pursuing Grand Challenges to set ambitious but achievable goals for AI;
funding rigorous evaluations of AI applications to measure their impact and cost-effectiveness;
and creating a policy, legal, and regulatory environment that allows innovation to flourish while protecting the public from harm.
Using AI in Government to Improve Services and Benefit the American People
One challenge in using AI to improve services is that the Federal Government’s capacity to foster and harness innovation in order to better serve the country varies widely across agencies. Some agencies are more focused on innovation, particularly those agencies with large R&D budgets, a workforce that includes many scientists and engineers, a culture of innovation and experimentation, and strong ongoing collaborations with private-sector innovators. Many also have organizations that are specifically tasked with supporting high-risk, high-return research (e.g., the advanced research projects agencies in the Departments of Defense and Energy, as well as the Intelligence Community), and fund R&D across the full range from basic research to advanced development. Other agencies like the NSF have research and development as their primary mission.
But some agencies, particularly those charged with reducing poverty and increasing economic and social mobility, have more modest levels of relevant capabilities, resources, and expertise.30 For example, while the National Institutes of Health (NIH) has an R&D budget of more than $30 billion, the Department of Labor’s R&D budget is only $14 million. This limits the Department of Labor’s capacity to explore applications of AI, such as applying AI-based “digital tutor” technology to increase the skills and incomes of non-college educated workers.
DARPA’s “Education Dominance” program serves as an example of AI’s potential to fulfill and
accelerate agency priorities. DARPA, intending to reduce from years to months the time required for new Navy recruits to become experts in technical skills, now sponsors the development of a digital tutor that uses AI to model the interaction between an expert and a novice. An evaluation of the digital tutor program concluded that Navy recruits using the digital tutor to become IT systems administrators
frequently outperform Navy experts with 7-10 years of experience in both written tests of knowledge and real-world problem solving.31
Preliminary evidence based on digital tutor pilot projects also suggests that workers who have completed a training program that uses the digital tutor are more likely to get a high-tech job that dramatically
30 Thomas Kalil, “A Broader Vision for Government Research,” Issues in Science and Technology, 2003.
31 “Winning the Education Future: The Role of ARPA-ED,” The U.S. Department of Education, March 8 2011, https://www.whitehouse.gov/sites/default/files/microsites/ostp/arpa-ed-factsheet.pdf.
16
increases their incomes.32 These wage increases appear to be much larger than the impacts of current workforce development programs.33 Ideally, these results would be confirmed with independently
conducted, randomized, controlled trials. Currently, the cost of developing digital tutors is high, and there is no repeatable methodology for developing effective digital tutors. Research that enables the emergence of an industry that uses AI approaches such as digital tutors could potentially help workers acquire in- demand skills.
32 The President’s Council of Advisors on Science and Technology, letter to the President, September 2014,
https://www.whitehouse.gov/sites/default/files/microsites/ostp/PCAST/pcast_workforce_edit_report_sept_2014.pdf.
33 J.D. Fletcher, “Digital Tutoring in Information Systems Technology for Veterans: Data Report,” The Institute for Defense Analysis, September 2014.
Recommendation 3: The Federal Government should explore ways to improve the capacity of key agencies to apply AI to their missions. For example, Federal agencies should explore the potential to create DARPA-like organizations to support high-risk, high-reward AI research and its application, much as the Department of Education has done through its proposal to create an “ARPA-ED,” to support R&D to determine whether AI and other technologies could significantly improve student learning outcomes.
Recommendation 4: The NSTC MLAI subcommittee should develop a community of practice for AI practitioners across government. Agencies should work together to develop and share standards and best practices around the use of AI in government operations. Agencies should ensure that Federal employee training programs include relevant AI opportunities.
17
AI and Regulation
AI has applications in many products, such as cars and aircraft, which are subject to regulation designed to protect the public from harm and ensure fairness in economic competition. How will the incorporation of AI into these products affect the relevant regulatory approaches? In general, the approach to regulation of AI-enabled products to protect public safety should be informed by assessment of the aspects of risk that the addition of AI may reduce, alongside the aspects of risk that it may increase. If a risk falls within the bounds of an existing regulatory regime, moreover, the policy discussion should start by considering whether the existing regulations already adequately address the risk, or whether they need to be adapted to the addition of AI. Also, where regulatory responses to the addition of AI threaten to increase the cost of compliance or slow the development or adoption of beneficial innovations, policymakers should consider how those responses could be adjusted to lower costs and barriers to innovation without adversely impacting safety or market fairness.
The general consensus of the RFI commenters was that broad regulation of AI research or practice would be inadvisable at this time.34 Instead, commenters said that the goals and structure of existing regulations were sufficient, and commenters called for existing regulation to be adapted as necessary to account for the effects of AI. For example, commenters suggested that motor vehicle regulation should evolve to account for the anticipated arrival of autonomous vehicles, and that the necessary evolution could be carried out within the current structure of vehicle safety regulation. In doing so, agencies must remain mindful of the fundamental purposes and goals of regulation to safeguard the public good, while creating space for innovation and growth in AI.
Effective regulation of technologies such as AI requires agencies to have in-house technical expertise to help guide regulatory decision-making. The need for senior-level expert participation exists at regulating departments and agencies, and at all stages of the regulatory process. A range of personnel assignment and exchange models (e.g. hiring authorities) can be used to develop a Federal workforce with more diverse perspectives on the current state of technological development. One example of such an authority is the Intergovernmental Personnel Act (IPA) Mobility Program, which provides for the temporary assignment of personnel between the Federal Government and state and local governments, colleges and universities, Indian tribal governments, federally funded research and development centers, and other eligible organizations. If used strategically, the IPA program can help agencies meet their needs for hard- to-fill positions and increase their ability to hire candidates from diverse technical backgrounds. Federal employees serving in IPA assignments can serve as both recruiters and ambassadors for the Federal workforce. For example, agency staff sent to colleges and universities as instructors can inspire students to consider Federal employment. Likewise, programs that rotate employees through different jobs and sectors can help government employees gain knowledge and experience to inform regulation and policy, especially as it relates to emergent technologies like AI.
34 Ed Felten and Terah Lyons, “Public Input and Next Steps on the Future of Artificial Intelligence.”
